Aspect Based Sentiment Analysis For United States Of America Airlines

2 2.1 Literature Review of Aspect Based Sentiment Analysis for US Airlines (2010-2017) Introduction In recent years, marketing styles and approaches have significantly changed (Mangold and Faulds; 2009). A lot of organizations analyze and rely on customers’ opinions in order to get valuable insights about their services and goods from different online platforms like social media, blogs or online review portals. Consumers also utilize these channels not only to express their opinions but also to get the views of other consumers about the commodity or services. Hence, customers’ opinions play a very important aspect in this buyer-seller relationship. With the increase in digital marketing and opinion sharing on virtual networks, many researchers have been attracted towards opinion mining and text analysis in the last decade. Opinion mining is not only profitable for the reviewer but also it is important for a service provider to find the major strengths and drawbacks in their business models by having a direct view from their customers. The work of Jindal and Liu (2006) defines opinions in majorly two forms, Regular opinion and Comparative opinion. In a Regular opinion, the person conveys their perspective in the form of a positive, negative or neutral comment about a product. Whereas, a Comparative opinion can be defined as when a person conveys a comparison between products while expressing his/her opinion. For example- Alaska Airline customers satisfaction is better than that of Jetblue Airlines. Here there is a comparison of customer satisfaction between the airline carriers. Further sentiment analysis can be derived from a deeper analysis of other areas. The next section is reflecting one of the most important fields of sentiment analysis, i.e. aspect based analysis. Aspect based sentiment analysis is important to find the polarity of text in the sentiment analysis. 2.2 A critical review of Aspect-based sentiment analysis Understanding opinions is a vital part in sentiment analysis but an opinion only exposes the polarity of a text in a sentiment analysis. Collomb et al. (2014) classify sentiment analysis into three levels which are document level, sentence level and word/feature level or can be said aspect level. In the document Level, the document is consider as a whole and taken as either positive or negative but the only drawback of this is that we don’t know which part of the document is positive or negative. Whereas, in the sentiment and Aspect level, the opinion is retrieved at a finite level and we can analyze which part or aspect the customer liked or disliked. This research will be carried out on aspect level of sentiment analysis for both regular and comparative opinions. 5

Figure 2: Aspect based sentiment analysis approaches (Schouten and Frasincar; 2016) 2.3 Identified Research Gaps in Airlines Industry In recent years, sentiment analysis has been performed in various domains from analyzing movie reviews to stock marketing analysis but little research has been done in the airline industry. Wu and Liao (2014), looked at the leading and lagging parameters for 38 airline companies from the annual and business reports generated from data envelopment analysis(DEA). Similar to that, Hannigan et al. (2015) analyzed the relationship between the different performance factors in the United States of America based airlines from 19962011 from the annual reports provided by DEA. In the results, they found that there is a positive relationship between the price and performance (in terms of stock market shares) over the years using regression models and found a negative relationship with the quality of the services. The authors discuss future work in terms of analyzing other data resources that could bring different insights to the research. Sultan and Simpson Jr (2000) gathered surveys from US and European travellers and built a SERVQUAL model from customer expectations and perceptions towards an airline’s performance. In the model, the reviews were segregated into five factors i.e Reliability, Assurance, Tangibles, Empathy and Responsiveness. Similarly, Min and Min (2015) noticed 18 factors that can provide a benchmark in the airline services. Extending the work of Min and Min (2015), Anitsal et al. (2017) analyze the sentiment of passengers for top ten U.S airlines from Skytrax. The researchers generated a sentiment score and word cloud of the reviews for each airline using the tool Semantria. Skytrax is the leading global airline consultancy firm which conducts annual surveys and conducts star based analysis under a global airline rating program (1-5) (Yakut et al.; 2015). The program gives feedback for more than 670 airlines performance globally (Pérezgonzález and Gilbey; 2011). Out of many 6

researchers Jansen et al. (2009) mention free accessible, experienced auditing, the unique ranking system and competent textual reviews as a major backbone of Skytrax rating system. In 2009, Skytrax faced controversies as mentioned by Jansen et al. (2009), out of all the airline open for review only 25 percent were scored that do not portray a real picture of the Airline industry. Hence, for a better textual analysis, heterogeneous sources are needed to cover the full range and all aspects of a particular industry. Safko (2010) mentions that social media is a reliable source of data for sentiment analysis as the data on such media keeps on updating. However, Ivanscenko (2016) points data collection and textual analysis for the random data makes the process more difficult, but Yee Liau and Pei Tan (2014a) state text mining as a major contributor in the findings of brand awareness, loyalty and recognition in the aviation industry. Li (2017) has focused on the gaps in sentiment analysis in order to fulfil criteria like reliability, discriminate validity, and external validity for airline review site SKYTRAX. The study reveals how research on SKYTRAX using numerical rating is non-reliable to judge the performance of any airline. According to Kaur and Duhan (2015), converting textual information into a meaningful analysis faces six critical issues: negation handling, domain generalization, pronoun resolution, language generalization, related knowledge and mapping slangs. Li (2017) has done an exceptional research in handling these hurdles to an acceptable degree, but there is still some room for errors while dealing fake or spam reviews.This research is helpful for proceeding our research towards the aspect based analysis as Document-based analysis lacks for extracting features from the content. As Li (2017) has researched strictly around airline industry deriving data from Skytrax and twitter, there are some points which lead this research under few doubts. External validity is an important part of the Skytrax global rating system and this feature has not been tested by the researchers. According to Cachia et al. (2007) organising unstructured data in a systematical way can be an issue as data is derived from many sources like newspapers, blogs, social media etc., which can lead to ambiguity and repetition of data. Text analysis faces the problem of normalisation which should be handled carefully, and this feature is very rare to come across in the existing research. The airline industry is being researched and analysis is done by many researchers, out of those Li (2017) performed most efficient manner of sentiment analysis. But author follows document based analysis which can further be improved by aspect based analysis on heterogeneous data source. Sankaranarayanan and Rathod (2017) classify the airline ratings for low-cost Indian airlines. The research was done on three low-cost Indian airlines from the review ratings generated from the Trip Advisor website using Expectation Maximization clustering techniques. Further, for prediction accuracy, a comparison was drawn among various machine learning techniques like Kmean, Naive Bayes and logistic model trees. The authors suggest a future work in considering text mining for the reviews in order to find the reason behind the rating of the reviews. There were other limitations in the research like missing data values and not considering other important aspects like baggage issues and seat comfort in the dataset that should be addressed. With the rising competitive conditions in the low-cost carrier (LCC) airline industry, a notable work has been accomplished in the past. O’Connell and Williams (2005) conducted a passenger survey for two Malaysian Airlines (Malaysia Airline and AirAsia) and found very interesting insights about the passengers thinking for choosing an airline. In the research, it is observed that the young age group of passengers was highly attracted towards the LCC airlines because 7

of their ticket cost whereas other segments of passengers chose their airlines on the basis of the staff services. It is found that the Air Asia airline staff have a very high productivity level which results in excellent customer satisfaction. Similarly, in the research of Saha and Theingi (2009) on Thailand Airlines, they found customer satisfaction as the most important factor in the buying behaviour of the customers. Dobruszkes (2006)researched the European LCC and noted a significant decrease in the cost of airline fares and an increase in air traffic. The researcher suggested that in order to be competitive in the market of LCCs, organizations should pay significant importance to their customers welfare.This proves that Airline industry need features to be focused on which can be derived from the reviews of the customer. There has been ample study conducted towards the LCCs but solely through employing passenger surveys. In contrast, Yee Liau and Pei Tan (2014b) analysed customer opinions through the microblogging site Twitter for LCCs in Malaysia; Twitter tweets were collected for two and a half months for the Airlines. Since most of the tweets were in the local Malaya Language, hence, a Malay lexicon was built for sentiment analysis and further the tweets were segregated into four different clusters, i.e. ticket promotions, flight cancellations, delays and customer satisfaction. Clustering is performed by using Kmeans and spherical Kmean algorithms. In the results, there was an overall satisfaction with the services provided by Malaysian LCCs. Apart from the research in LCC airlines, there has been some other research done toward the airline industry. Dharmavaram Sreenivasan et al. (2012) analysed sentiment polarity for three airline groups from the tweets extracted for a period of four months using lexicon analysis. Miner (2012) conducted research on predicting the airline ranking for US based airlines. The research follows the Variational Expectation Maximization (VEM) technique to determine the airline ranking and the Correlated Topics Models (CTM) for extracting the sentiment topics. Also, Baumgarten et al. (2014) predicted the delay in flights by analysing the departure and arrival times of the US-based flights. 2.4 Conclusion To the best of the candidates knowledge, it is clear based on the reviewed literature that there is an inadequate amount of research done in determining the Aspect level of consumer sentiment in the airline industry. Most of the research focuses either on finding the sentiment polarity of the consumers experience with airline services or focuses on determining the rating or delay time in the airlines. Also, in order to get proper insights into customer behaviour towards airlines, utilizing surveys or Twitter feeds for two to three months is not adequate; hence, there is a need for heterogeneous data sources. Based on the identified gap there is a need to develop an Aspect-based decision support system for the US Airline Industry. 3 3.1 Methodology Introduction The Methodology section comprises, the process flow of the research along with the discussion of the methodology used in the research and a brief explanation of each part of the methodology. Lastly, there is a conclusion of this section. 8

3.2 Modified KDD Methodology The process of extracting essential insights from the data to provide a better decisions support is known as Data Mining (Berry and Linoff; 1997). Methodologies such as knowledge discovery and data mining (KDD) and Cross-Industry Standard Process for Data Mining (CRISP-DM) are widely used for data mining (Azevedo and Santos; 2008). The CRISP-DM methodology was introduced in 1996 and follows a six-stage process flow (Chapman et al.; 2000). Since in this research, stages like business understanding and development cannot be utilised, hence CRISP-DM methodology does not fit into this research. Furthermore, This research will follow a modified KDD methodology where a few stages are modified e.g. implementation stage modified to Aspect and sentiment detection stage and additional attributes are added. Hence, the modified KDD methodology fits best for the research. The proposed methodology is shown in Figure 3. Figure 3: Modified Methodology 3.3 Data Collection This is the first step of the methodology. The data is extracted from various sources like Twitter tweets and online reviews from Skytrax (2014-2017) for the top 10 US-based airlines. The main purpose of collecting the data is to find out the leading and lagging performance parameters among the airlines. 3.4 Data Pre-processing After the data collection, the text data is cleaned and cleansed. This phase of the methodology is the most important phase; in order to build a robust, stable and efficient system, the data must be properly cleaned. Each review and tweet was converted to a sentence after the data pre-processing stage for further analysis. 3.5 Aspect and Sentiment Detection Once the data is cleaned and the reviews and tweets were converted to cleaned sentences, the sentiment was extracted from each sentence and a polarity was assigned. Similarly, Aspects were detected for each sentence and a separate column was made for each aspect 9

with its polarity score. Finally, the aspects and sentiment of the sentences were stored in a tabular form for each airline. 3.6 Data Mining Statistical operations like correlation and linear regressions were performed on the aspects and polarity score of the reviews for the airlines. Hence, these statistical operations provide a better insight into the data and help by understanding the relationships among different factors in the dataset. 3.7 Interpretation, Evaluations and Visualization The evaluation of the research is based on the Precision, recall and F-score values. Further, various machine learning algorithms were performed in order to find the best-fit algorithms for the system. Last but not the least, the results were visualized using Tableau and a case study was formed in order to answer the proposed research questions. 3.8 Conclusion Hence, for this research a modified KDD approach is implemented as it fits best for the research. Furthermore, a detail explanation of the steps are discussed in the following implementation section. 4 4.1 Implementation Architecture Design The proposed research follows the architecture design shown in Figure 4. It highlights the tools and technologies that are used in the research. The proposed design has a good inter connectivity with the stages which corresponds with the modified KDD methodology. First part of the design is the Data Persistent layer which deals with the data sources, Data processing, and Databases. This layer is connected to the Business project requirement layer which provides the solution to the Business problems and the final stage, Client presentation layer, where the client and data analyst interact and the reports and dashboards are presented. 10

Figure 4: Architecture Diagram 4.2 Process flow The process flow diagram for the research is shown in Figure 5. The data for the research was taken from two sources, Twitter and online review site Skytrax. For this task, the twitterR package in Rstudio was used (Gentry; 2012) to collect Twitter tweets for the Airlines. API keys and a secret key was generated for authentication access to Twitter. Once the configuration was setup in R, Twitter tweets for airlines were collected for 18 days and a total of 60,445 tweets were collected for the research. The data consists of tweets, airline name, and airline code. Once the data were gathered it was extracted to a comma separated file (CSV) file and stored in the system and external storage device. For the Skytrax data source, the data was scrapped using Google Chrome Plugin ”Data Miner” which extracts data automatically.The tool extracted the airline reviews, airline name, timing, location, reviewer name, and their review ID. The reviews were collected from 2014 to November 2017 and each airline review was collected separately and extracted to a CSV file. All the CSV files from Skytrax reviews were consolidated in a single CSV file. In the end, both the data sources from Twitter and Skytrax were consolidated into a single dataset. 11

Figure 5: Implementation Process flow Diagram After arranging the consolidated data-set, unnecessary observations like author names, locations, and review Ids were removed in order to maintain ethical privacy issues. Further, the data was moved to a data frame and was converted into a corpus in order to perform various text processing methods which includes removing the punctuation, converting the text into lower case, removing white-spaces, stemming and removing empty rows. Also, additional customized stop-words were added in order to remove the stopwords using ’tm’ package in R. Figure 6 illustrates the data pre-processing. Figure 6: Data Pre-Processing 12

The gathered data was then interpreted as 1-gram, 2-gram and 3-gram manually by inspecting their frequency using Wordle, then passed to the part of speech (POS) tagging part using the openNLP package in R. Where each word in the sentence is tokenized and tokenized words were split into new sentences. POS tags were extracted as it provides noun, pronoun, adjective etc. for each word. As there was no proper dataset available for training data, th

Southwest Airlines, Spirit Airlines, United Airlines and US Airways. . to have a good competitive edge from their competitors. To the best of the candidate's knowledge, there has not been much research done on . "Can aspectbased sentiment analysis of 10 US Airlines namely Alaska Airlines, America Airlines, Delta Airlines, Hawaiian .